Extending

The other direction: GoTo Agents as an MCP server.

Every agent tool attaches MCP servers. GoTo Agents also is one: point Claude Code, Claude Desktop, or any MCP client at goto_agents_mcp.py and that client can open sessions, message your agents, and run your tasks — your studio becomes a capability in tools you already use.

Setup

{
  "mcpServers": {
    "goto-agents": {
      "command": "python",
      "args": ["C:/github/goto-agents/goto_agents_mcp.py"]
    }
  }
}

The connector is an adapter over the Python client, so it needs a running GoTo Agents server to talk to.

What a client gets

  • Orientation — get_server_info returns server status, profiles, and workspaces, with a suggested next step baked into the response so LLM clients can self-navigate.
  • Sessions and messages — new_session(profile_id, workspace_name, agent_name) and send_message: a Claude Code session can literally converse with your co-writer agent.
  • Tasks — gta_run_task and gta_resume_task for structured JSON work orders. See Tasks.
  • Resources — task definitions are exposed as MCP resources, so a client can read what a task expects before invoking it.
  • Project context — set_project_directory points the session at a folder, same as the Project tab in chat.

Why this matters more than it sounds

Pair it with AIR and the relationship becomes genuinely two-way: the Claude Code session driving your agents can also serve their requests — lending its web access, local filesystem, and credentials when a GTA agent asks. Client and server stop being roles and become two collaborators with different capabilities, which is the product thesis restated in protocol form.

Where it lives

  • goto_agents_mcp.py — the MCP server (FastMCP tools + resources).
  • gotoagents/client.py — the client it adapts.
  • help/mcp.md in the repo — in-app operational notes.